AI Engineer Intern working across product software, backend reliability, cloud tooling, model evaluation, and coding-agent systems.
annotation / review software
Built and maintained a multi-user computer-vision annotation and review platform on Cloudflare Workers, Supabase/Postgres/Auth, and R2, with reviewer queues, edit locks, minor-version history, resumable imports, reproducible training exports, aggregate dashboards, and concurrent-save protections.
correctness / performance
Tracked issues around image navigation, saves, locks, duplicate metadata, prefetching, dashboard cost, access control, and observability with layer-by-layer measurement, request IDs, failure classes, targeted fixes, and path-level verification.
shared gpu coordination
Built a shared EC2/GPU coordination service with a realtime fleet dashboard, Python CLI, stable identity, token-scoped actions, waiting queues, and declared usage so people and coding agents could avoid stepping on each other.
agent tooling
Built shared-context tooling for coding-agent work with append-only records, namespaces, verified handoffs, secret rejection, REST/MCP surfaces, export/restore, and recovery paths for long-running tasks where memory drift creates real bugs.
document software
Worked on Android bulk answer-sheet capture with CameraX preview, analysis, and capture; automatic page detection; ONNX document-corner inference; quality gates; page-turn/re-arm logic; batch resume; raw JPEG storage; and JSON/EXIF capture metadata.
training / evaluation
Used reviewed datasets, detector training and comparison, source-separated evaluation, visual checks, leakage gates, frozen tests, and comparative benchmarks to guide model promotion and stop decisions.
Four backend and agent systems spanning realtime audio, research automation, data analysis, and portable context.
The current path uses authenticated WebSockets, separate transcription and speaker lanes, a sequence-derived session clock, append-only Postgres revisions, deterministic chunks, Postgres lexical search, Qdrant dense retrieval, and SSE cited chat.
A LangGraph fan-out/fan-in sends web research into Qdrant, runs parallel job and GitHub branches, lets a curator rank targets, preserves the human pause in SQLite, and keeps the isolated ReAct GitHub path behind tool and input guardrails.
The system uses sqlglot recursive classification across reads, writes, DDL, CTEs, and subqueries; fail-safe policy decisions; least-privilege reads; PII redaction; approval-gated writes; hash-chained audit; and SDK, REST-SSE, MCP, and LangGraph surfaces.
It uses append-only records, schema and policy checks, content hashes, namespaces, secret rejection, supersede/relocate/split/merge events, export/restore, and CLI, REST, MCP, and skill-pack surfaces for verified handoffs.
Co-developed · backend/security by me
Wombat is an API-key vault for teams that need scoped access without exposing raw provider secrets. It handles ownership checks, encrypted storage, rate limits, validation, and documented API contracts.
Co-developed
Saturday is a voice search tool for codebases. It indexes a repository, connects a voice assistant to a local webhook, and lets a developer ask spoken questions while working.
Solo
Reclaym is a UPI AutoPay recovery engine for failed subscription payments. It receives payment events, deduplicates retries, applies deterministic recovery policy, and reconciles payment-state changes.
Co-developed
Headless Ghidra MCP is a native inspection bridge for reverse-engineering workflows. It exposes Ghidra analysis, decompilation, imports, exports, call graphs, JSON output, pagination, caching, and safer path handling.
Tools and engineering practices used across backend products, cloud workflows, reliability work, and applied-AI systems.
Service boundaries, authentication, realtime delivery, and documented contracts.
Durable state, queues, retrieval, and consistency under concurrent work.
Production diagnosis, observability, automated tests, and evidence-driven evaluation.
Edge and cloud runtimes, object storage, deployment, and operator tooling.
Retrieval, agents, speech and document pipelines, and model evaluation inside products.
Coding activity.
I am most useful where software has to hold up: APIs, databases, reliability work, evals, product experiments, internal tools, and applied-AI systems with clear boundaries.